Routers Are a New AI Primitive

AI model performance varies substantially across professional tasks such as finance, coding, law, and tax, making reliance on a single frontier model potentially costly and less effective than selecting models by workload. Vals argues that AI routers, which choose among models for individual requests or steps in an agent workflow based on quality, cost, and latency, can address this trade-off, with a theoretical oracle router outperforming the best single model by 14 points on the Vals Index while costing roughly 40% less. Existing router benchmarks often rely on isolated, public, and increasingly saturated questions, which fail to capture the compounding effects of routing choices during long-running, tool-using agent tasks, including errors, context growth, prompt-cache losses, and recovery costs. Vals proposes an independent evaluation framework using its continuously updated Index of eight mostly private, expert-designed benchmarks across professional domains, running routers live within identical agent harnesses and comparing outcomes, cost, cache usage, and completion time against the best and cheapest models in each router’s pool and an oracle reference. The organization plans to publish a public router leaderboard to assess whether routing systems can deliver practical gains beyond individual models.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.